You bought the accounts. Nobody uses them.

Somebody bought everyone a subscription in January. Two people tried it, decided it was a toy, and it's been renewing quietly ever since. That's not a failure of your people and it isn't a failure of AI. It's what happens when the setup step gets skipped, and the setup step is most of the work.

I've spent about 25 years in manufacturing and owned and ran a machine shop for the better part of a decade. I now build these tools for a living, and about ten of them run every day inside a real precision manufacturing operation. Here's what I've watched go wrong.

1. It doesn't know anything about your business

This is the big one and it explains most of the rest.

An AI account out of the box knows an enormous amount about the world and precisely nothing about you. It has never seen your part numbers, your customers, your vendors, your procedures, your routings, or the way your quotes actually get built. So the first real question anybody asks it gets a generic answer, the person decides it doesn't understand their work, and they stop asking.

They're right, by the way. It didn't understand their work. Nobody had told it anything.

The fix isn't a better prompt, it's loading your operation into it: your procedures, your specs, your history, connected to the systems that hold your real data. That's unglamorous work and it's the whole difference between a tool and a novelty. There's more on how that's done on building an AI knowledge base that knows your business.

2. It was pointed at the wrong half of the job

AI is genuinely good at reading. Messy documents, acknowledgments laid out differently by every vendor, a request with the quantities buried in a paragraph, twenty years of procedures nobody can find.

What it will not do is be exactly right every single time. So the moment somebody points it at a price, a quantity, or a record going into the ERP, it gets one wrong, and the trust is gone for good. Not unreasonably.

Give it the reading and give ordinary code the arithmetic and this problem disappears. It's the single most useful distinction to get straight before spending money, and it has its own page: AI is a good reader, it's a bad accountant.

3. Nobody had a specific job for it to do

"Use AI to be more efficient" is not an assignment. It's a hope.

The rollouts that stick start from a job somebody actually has to do on a Tuesday. Read this acknowledgment and tell me if the vendor changed the date. Find me what our procedure says about this material. Draft the response to this customer complaint. Narrow, repeated, and obviously useful the first time it works.

The ones that don't stick start with everyone getting a login and a link to a training video.

4. The people who could tell you what to automate were never asked

The person who knows where the week goes is the one doing the typing, and they're usually the last one in the room. So the project targets whatever was visible from the office, which is rarely the thing eating the most hours.

It also creates a quieter problem. If the first anybody hears about AI is that it's arriving, the honest read from the floor is that it's here to replace them. Nobody then volunteers what the boring parts of their job are, which is exactly the information the project needed. I've written separately on whether AI is coming for the skilled trades, and the short version is no, but not saying so out loud costs you the cooperation you need.

5. It was treated as a purchase instead of an implementation

Software you buy is done when it's installed. This isn't that. The accounts are the cheapest part of it. What makes it useful is the setup, and skipping the setup while paying for the accounts is how a business ends up with a subscription nobody opens and a conclusion that AI doesn't work in manufacturing.

The uncomfortable one, which is really about data

Sometimes it fails because the answer it would have to give is sitting in a place nothing can read. Job history in a system that won't export cleanly. Procedures in a binder. What a job actually cost, known accurately only after month end. Prices that live in somebody's head.

None of that is an AI problem and no amount of AI budget fixes it. It's worth finding out early, because it changes what's worth doing first. If the numbers aren't trusted, that's the thing to fix, and I've written up why job numbers usually aren't trusted separately.

What working looks like instead

Small, specific, and against real data on day one.

Pick one job that happens many times a week and has a checkable right answer. Load what the AI needs to know about how you do it. Connect it to whatever system holds the truth. Run it against real work, not a demo, and watch what breaks.

It will break. Almost every tool I've built broke the first time we ran it against real data, because real orders have exceptions nobody describes to you up front. Finding those is the work, not a sign the thing was a bad idea. Anyone who tells you their rollout went in clean has either not done one or isn't telling you.

Common questions

So was buying the accounts a mistake?

No. The accounts are fine and they're the cheap part. What's missing is the setup that makes them know your operation. That work can be done on top of what you already bought.

Which AI should we be using?

Whichever fits. I'm not tied to one vendor and I'd rather pick on fit than on what's easiest for me to sell. More on how that's set up on AI for manufacturers.

Is our data safe in these things?

It depends entirely on which product and which plan, and the differences are real. That question deserves a straight answer rather than a reassurance, so it has its own page: is ChatGPT safe for company data.

How do we tell where we actually stand?

The AI readiness assessment walks through your own operation and gives you a report at the end. It's free, nobody has to call you, and it'll tell you if the honest answer is that you're not ready yet.

Related reading. If the pressure behind all this is that you can't fill the office roles, that's the page that lists what to automate first. If the worry is what happens when a key person retires, capturing what only lives in one head is the same problem from the other end.